NewsResearch
Researchers map how world models in robots, cars and agents can be attacked, and why their confident predictions are a risk
World models, the learned simulators behind JEPA and other systems that predict what happens next, are moving into robots, autonomous vehicles and agents, and researchers are mapping how they fail. In April, Manoj Parmar's survey "Safety, Security, and Cognitive Risks in World Models" described adversaries corrupting training data, poisoning latent representations and exploiting compounding prediction errors; found that agents able to simulate the consequences of their own actions are more prone to goal misgeneralisation, deceptive alignment and reward hacking; and warned that authoritative predictions foster automation bias that operators lack the tools to audit. It extends MITRE ATLAS and the OWASP LLM Top 10 to the world-model stack. In July, a lifecycle review by Fazhong Liu and colleagues, which lists Meta's V-JEPA 2 among representative world models, added fooled sensors (adversarial patches, LiDAR spoofing), prompt injection and typographic attacks placed in the environment, and attacks that make a planner prefer a dangerous future, and named a "predictive safety illusion": when a world model is used as a safety checker, a confident but wrong prediction can make an unsafe action look certified. A September survey in Advanced Engineering Informatics argues that JEPA may resist pixel-level attacks better than today's driving systems.
Why it matters
A world model's mistakes become physical actions, not wrong sentences. If one is used to decide whether an action is safe, give it an independent check, know where its training data came from, and do not trust its sensors blindly.